TL;DR
- GPT-6.2 Astra has a score of 62 on the Intelligence Index (max and xhigh settings) and is 3rd of 96 models.
- xhigh is similar to max in 54% of the time per task (7.5 vs 13.9 minutes).
- The price is $8.00 per 1M input tokens and $40.00 per 1M output tokens, four times the median for similar models.
- Max waits 209.80 seconds for its first answer token. Medium waits 3.67 seconds.
- It can take text and images as input and output text, with a 1M-token context window.
GPT-6.2 Astra outperforms the Intelligence Index at 62 and is the 3rd model of 96. It's also less fast and more expensive than the median. The benchmark scores, the five effort levels, speed, token usage, pricing and best uses are discussed in this guide.
GPT-6.2 Astra benchmarks indicate that the model is faster and more intelligent than most other models, but it is more expensive. It was released by OpenAI on November 17, 2026. It ranks highest on the Artificial Analysis Intelligence Index among similar models, with a score of 62 compared with the median of 43. This places it 3rd out of 96. However, the cost is high: $8.00 per 1M input tokens and $40.00 per 1M output tokens. This guide outlines each figure so you can select the appropriate level of effort.
What GPT-6.2 Astra Is
GPT-6.2 Astra is OpenAI's in-house reasoning model released on November 17, 2026. It can take text and image input and provide text output. It has a 1M-token context window, or approximately 1,500 A4 sheets of 12-point Arial (100% scale).
The number of parameters has not been revealed by OpenAI. There's no public set of weights. The model is available via one API provider: OpenAI.
| Spec | Detail |
|---|---|
| Creator | OpenAI |
| Release date | November 17, 2026 |
| Model type | Proprietary reasoning model |
| Input | Text and image |
| Output | Text |
| Context window | 1M tokens (about 1,500 A4 pages) |
| Effort levels | Max, xhigh, high, medium, low |
| API providers | 1 (OpenAI) |
Simply put, the model can read text and images, write text, and accept a very long document in a single request. You select the difficulty level it has to think by selecting an effort level.
How GPT-6.2 Astra Compares Inside OpenAI’s Lineup
The Intelligence Index score of the GPT-6.2 Astra is the highest among all OpenAI models listed on its Artificial Analysis page. At 62, it beats GPT-6.2 Sol’s top score of 57 by 5 points, GPT-6.1 Terra’s 51 by 11 points, and GPT-6 Luna’s 28 by 34 points. At medium effort, Astra also scores 57, matching Sol's maximum score.
Artificial Analysis records five versions of Sol, six of Terra and six of Luna. This comparison is based solely on intelligence, as the page text provides only the top intelligence rating for each. Before you decide on a model, visit each model's page to check its speed and price.
Also Read: GPT-6.2 Sol vs GPT-6.1 Sol for how the cheaper tier compares with the model before it.
GPT-6.2 Astra Benchmarks Start With a 62 on the Intelligence Index
At its maximum and xhigh settings, GPT-6.2 Astra achieves scores of 62 on the Artificial Analysis Intelligence Index (AAII), placing it 3rd out of 96 models. The median of similar models is 43. Even with the lowest setting (46), exceeds the median.
The index is a composite score. The numbers indicate the strength of the results on the tests that the number encompasses. Astra is also at the top of its own family in this index. OpenAI’s GPT-6.2 Sol tops out at 57, GPT-6.1 Terra at 51, and GPT-6 Luna at 28.
| Effort level | Intelligence Index score |
|---|---|
| Max | 62 |
| Xhigh | 62 |
| High | 60 |
| Medium | 57 |
| Low | 46 |
| Median, comparable models | 43 |
The more effort they put into each step, the higher the score. The largest drop is from medium to low: 11 points. The steps above medium are "low cost" (only 3 points or less each).
The Index Combines Nine Tests of Real Work
The Intelligence Index (v5.0.1) combines nine assessments of agentic work, coding, science and knowledge. 5 out of 9 are agentic; this requires the model to perform actions, not just respond to questions. This is more like a work test than a quiz.
| Evaluation | What it measures |
|---|---|
| AA-Briefcase v1.1 | Agentic knowledge work |
| AutomationBench-AA | Agentic SaaS workflows |
| Terminal-Bench 4.0 | Agentic coding and terminal use |
| FrontierCode-AA 2.0 | Agentic software engineering in open-source repositories |
| AA-Omniscience v1.1 | Knowledge |
| GDP.pdf | Professional document reasoning |
| HLE-Diamond | Reasoning and knowledge |
| CritPt v2 (Hard Subset) | Physics reasoning |
| Terminal-Bench-Science 0.1 | Agentic scientific research workflows in a terminal |
Office processes, software engineering, science and hard reasoning all contribute to the total. The composite score is reported in this guide. Per-evaluation charts are displayed on the model page in Artificial Analysis.
Effort Levels Change Score, Time, and Token Use
Medium is far behind Low. This contributes 11 points for 2.6 additional minutes per task.
Add a little over Medium to High. It gets 3 points for an additional 1.6 minutes per task.
Pick the level that fits the task, not the one with the biggest label.
| Effort level | Index score | Minutes per task | Output tokens (Index run) |
|---|---|---|---|
| Max | 62 | 13.9 | 65.8M |
| Xhigh | 62 | 7.4 | 27.1M |
| High | 60 | 6.1 | 20M |
| Medium | 57 | 4.5 | 12.4M |
| Low | 46 | 1.9 | 6.31M |
Three patterns stand out:
- Xhigh matches max. It uses about 41% of the output tokens (27.1M vs 65.8M) and 54% of the time per task.
- Medium beats low by a wide margin. It adds 11 points for 2.7 extra minutes per task.
- High adds a little over medium. It gains 3 points for 1.6 extra minutes per task.
The first step up in effort buys the most. The last step, from xhigh to max, adds no points on this index. Index parity does not guarantee parity on your own work, so test both on your tasks.
Speed: About 50 Tokens per Second, With a Long Wait at Max
GPT-6.2 Astra runs at around 50 tokens per second at max, xhigh, and high levels of effort, and slightly less at medium (49) and low (48) levels. The larger delay precedes the first word. The maximum response time to the first token is 209.80 seconds (approximately 3.3 minutes). The median for similar models is 12.92 seconds.
| Effort level | Output speed (tokens/sec) | Time to first answer token |
|---|---|---|
| Max | 49 | 209.80 s |
| Xhigh | 49 | 71.63 s |
| High | 49 | 29.08 s |
| Medium | 48 | 3.67 s |
| Low | 46 | 2.81 s |
The overall speed of the task is in the middle range. The maximum time Astra requires for Intelligence Index tasks is 13.9 minutes per task, approximately 15% more than the 12.1-minute median. Artificial Analysis ranks 55th out of 95 in performance.
Time to first answer token refers to the model's thinking time. It takes more than 3 minutes for max mode to respond. Medium: Answer within 4 seconds, which is appropriate for chat-style use. Slow: Answer in less than 4 seconds, which is appropriate for chat-style use.
Pricing: $8.00 in and $40.00 out per 1M Tokens
The price of GPT-6.2 Astra is $8.00 per 1 million input tokens and $4.00 per 1 million output tokens. Both are quadruple the median price of similar models. Model 65 ranks 65th out of 95 models by Artificial Analysis for cost.
Cached prompts come with a 95% discount, or approximately $0.40 per 1M cached input tokens. The blended price is $5.88/1M tokens with a ratio of cache hit:input:output of 7:2:1. The average cost of running one Intelligence Index task at max is $4.00.
| Price item | GPT-6.2 Astra | Median, comparable models |
|---|---|---|
| Input, per 1M tokens | $8.00 | $2.00 |
| Output, per 1M tokens | $40.00 | $10.00 |
| Cache discount | 95% | – |
| Blended, per 1M tokens (7:2:1) | $5.88 | – |
| Cost per Index task (max) | $4.00 | – |
The blended price is the same across all five effort levels. When you increase effort, you still pay more on your bill because you write more tokens at a higher level. The greatest input-side savings come from reusing long prompts via caching.
Also Read: GPT-6 Astra pricing for the earlier Astra’s rate card, since 6.2 Astra is priced 20% below it.
Token Use: Max Mode Writes 65.8M Output Tokens
During the Intelligence Index run, GPT-6.2 Astra (max) produced 65.8M output tokens, slightly above the median of 65.1 m. According to Artificial Analysis, this is “somewhat verbose.” The model read 2.16B input tokens, so the output is 3.051% of the input. It ranks 49 of 95 on the verbosity list.
The sooner the effort begins to diminish, the faster token use fades. Xhigh wrote 27.1M tokens, high 20M, medium 12.4M, and low 6.31 m. Low uses 10.4 times fewer tokens than Max. Costs are driven by the number of tokens, at $40.00 per 1M tokens.
How Speed, Cost and Tokens were Measured
Artificial Analysis calculates the time and cost per task based on measured token counts, speeds and costs. Time per task equals output tokens divided by output speed, with output speed weighted according to the proportion of the Intelligence Index each test provides. Cost per task is the sum of input, cache hit, cache write, reasoning and answer token prices, each weighted equally.
Average speeds and latencies are calculated over the past 5 days. The figures may change over time as Artificial Analysis runs its benchmark every six hours. Consider today's figures a snapshot, and review the model page before planning a budget.
Output speed includes reasoning tokens and covers the entire request (from the prompt to the last token). The max mode produces the most tokens and is therefore the slowest per task.
Capability Indexes: Group the Tests by Profession
Artificial Analysis also publishes six capability indexes that group the same tests into fields. Different skills matter to a legal team, and different skills matter to an engineering team, so these indexes can be used to match a model to a job.
| Capability index | Evaluations included |
|---|---|
| Finance & Accounting (5) | AA-Briefcase, AA-Omniscience, HLE-Diamond, AutomationBench-AA, GDP.pdf |
| Strategy & Ops (4) | AutomationBench-AA, AA-Briefcase, AA-Omniscience, GDP.pdf |
| Legal (5) | AA-Omniscience, AA-Briefcase, HLE-Diamond, GDP.pdf, AutomationBench-AA |
| Healthcare & Medical (5) | AA-Omniscience, AA-Briefcase, MLCR-AA, AutomationBench-AA, HLE-Diamond |
| Engineering (7) | AA-Omniscience, CritPt v2, Terminal-Bench-Science 0.1, HLE-Diamond, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Briefcase |
| Economics (4) | AA-Omniscience, AA-Briefcase, HLE-Diamond, GDP.pdf |
Each index implements tests that apply to its respective field. Astra's scores for these indexes are located on the Artificial Analysis model page.
Which Effort Level to Choose
Choose xhigh for the top score at lower time and token cost than max. Choose medium for interactive work and low for simple tasks. If you are primarily limited by price, skip Astra altogether, as each level is four times the median per token.
| If you need… | Choose | Why |
|---|---|---|
| The highest index score | Xhigh | Scores 62, same as max, in 7.5 minutes per task |
| A balance of score and time | High | Scores 60 in 6.1 minutes per task |
| Fast first answers | Medium | Scores 57; first answer token in 3.54 s |
| The lowest time and token use | Low | Scores 46, still above the 43 median; 1.9 minutes per task |
| A final check on hard tasks | Max | Test it only if xhigh falls short on your own work |
To build effort routing into a workflow, so routine steps run at low and hard ones at xhigh, an AI agent builder lets you describe which tasks get which setting and where a person reviews.
Where GPT-6.2 Astra Fits Best
GPT-6.2 Astra is designed to do hard, multi-step work where quality is more important than speed and price. Five of nine Intelligence Index tests reflect agentic work; Astra's composite score of 62 is based on those tests. It also incorporates a 1M-token context window, enabling it to consume lengthy material in a single request.
Three situations correspond to the data:
- Long documents: This window contains approximately 1,500 A4 pages, and the GDP.pdf test in the index focuses on professional document reasoning.
- Image-based tasks: It can also be used with pictures, screenshots, charts, and text.
- Engineering & research activities: This index contains Terminal-Bench 4.0, FrontierCode-AA 2.0 and Terminal-Bench-Science 0.1 for coding and scientific workflows in a terminal.
The composite is not meant to provide information about how well Astra performs on any given test. Test it out alone for a task before you decide to commit.
Also Read: Best AI models for coding for how Astra ranks against other coding models.
Limits to Know Before You Commit
GPT-6.2 Astra is more intelligent than fast, but more expensive than GPT-6.0 Turbo. It's a compromise that works well in production. Artificial Analysis ranks it 55th out of 95 for performance and 65th out of 95 for cost, which is below its 3rd ranking in intelligence.
Max mode has a 209.80-second delay before the first answer token, which is unsuitable for live chat.
- The price of all input and output tokens is four times the median.
- Only text output is available.
- Only OpenAI is listed, as no price comparison is possible.
- The model is proprietary; its weights are not public, and the number of model parameters is not disclosed.
Conclusion
GPT-6.2 Astra is a premium reasoning model for demanding, multi-step work. It ranks 3 of 96 on the Intelligence Index with a score of 62, while its 1M-token context and image input suit long documents and visual analysis. However, it is slower and substantially more expensive than comparable models, so choose xhigh for top intelligence at lower cost, medium for interactive tasks, and low for routine work. Reserve max for cases where xhigh does not deliver sufficient quality. If you’re weighing Astra against other flagships, see GPT-6.2 Astra vs Gemini 4 Carbon or GPT-6.2 Astra vs Claude Fable 5.2.
But you can switch to customisation for better, more tailored results. Emergent is an app builder that turns a plain-language description into a working application. You describe the tool or app you need, and its AI agents design, build, and deploy it, with room to adjust features through chat as your needs change.

Describe it. Build it. Launch it.
- End-to-End Creation
- No Code Needed
- Live in Minutes

Most AI app builders stop at prototypes. Emergent creates production-ready apps you can actually launch.
- Production-ready apps
- Web & mobile apps
- Deploy in minutes






